DYNAMIC BEHAVIOR OF SHORTEST PATH ROUTING ALGORITHMS FOR COMMUN -ETC(U)JUN 80 D P BERTSEKAS N0001-75-C-1183
UNCLASSIFIED LIDS-TH-1005 NL
'U UlllllllEllIIIIEIIIIIIIIEEEEIIIIIIIIIEIIEIIIIIEIIEI
1ll 12 8____o12 1
IIIQ
I N N I H
.A 11,.NA i . .Al . . .A.,
June, 1980 -/LIDS-TH-1005
Research Supported By:DARPA Graxt ONR-NO0014-7
W..C-1183.
00 OSP No. 82933
DYNAMIC BEHAVIOR OF SHORTEST PATH ROUTING
ALGORITHMS FOR COMMUNICATION NETWORKS
DTIC.;.,, ,.. ,,.,e C L - %'T E !Dimitri F. Setekas)
SAUGS W0j
A
/ D~~~ISTP1IT'O~ ~~tAArpxO1"Ied fo- d rd a e
l €Laboratory for Information and Decision SystemsC.. . MASSACHUSETTS INSTITUTE OF TECHNOLOGY, CAMBRIDGE, MASSACHUSETTS 02139
80 84 34
June, 1980 LIDS-TH-1005
Dynamic Behavior of Shortest Path Routing
Algorithms for Communication Networks,
by
I.)Dimitri P./Bertsekas
Department of Electrical Engineering and Computer Science
Laboratory for Information and Decision Systems
Massachusetts Institute of Techqology
Cambridge, Mass. 02139
//
/ - i
Abstract
Several proposed routing algorithms for store and forward communication
networks, including one currently in operation in the ARPANET, route
messages along shortest paths computed by using some set of link lengths.
When these lengths depend on current traffic conditions as they must in an
adaptive algorithm, dynamic behavior questions such as stability, convergence,
and speed of convergence are of interest. This paper is the first attempt
to analyze systematically these issues. It is shown that minimum queuing
delay path algorithms tend to exhibit violent oscillatory behavior in the
absence of a damping mechanism. The oscillations can be damped by means
of several types of schemes two of which are analyzed in this paper. In
the first scheme a constant bias is added to the queuing delay thereby
providing a preference towards paths with small number of links. In the
second scheme the effects of several past routings are averaged as for
example when the link lengths are computed and communicated asynchronously
throughout the network.
-!7
2
1. Introduction
A central operational problem of a communication network involves the
choice of routes used by messages to travel from origin to destination. It
is possible, of course, to choose a fixed route for each origin-destination
pair, but this precludes the possibility of adjusting routes to alleviate
congestion due to variations in average traffic conditions. For this
reason attention has focused on adaptive routi ig strategies
whereby congestion in the network is continuously monitored and routes between
origin-destination pairs are modified in real time so as to keep average
delay per message at a reasonable level. A routing scheme of this type was
implemented in the ARPANET in 1969 and attracted considerable attention.
The main idea in this scheme is to compute in real time an estimate of the
minimum average delay per message for each origin-destination pair and to
route messages along the current minimum estimated delay path. When this
scheme was first implemented, it was noticed that it is
pione to severe oscillations. This behaviour is due to the fact that
delay estimates used to choose routes are themselves affected by the route
choice with a feedback effect resulting. To remedy this situation it was
decided on heuristic grounds to introduce an additive factor, called bias,
to the estimated delay of each link, thereby building into the algorithm
a preference towards paths with small number of hops to the destination
[5] - [7]. This had a stabilizing effect albeit at the expense of considerable
loss of sensitivity to traffic congestion.
The implementation of the minimum delay path idea in the original
ARPANET algorithm had a number of flaws allowing, for example, the formation
of loops. For thj! reason alternative schemes based on the same idea were
3
studied, and a new algorithm called SPF has been developed and imple-
mented [11, [4], [11). The present paper is an outgrowth of the
author's participation in the design study of this algorithm during the summer
of 1978 at BBN, Inc. However, our analysis does not focus on the ARPANET
and the SPF algorithm in particular, but rather is geared towards understanding
the effect of feedback and the nature of the dynamic behaviour of shortest
path algorithms where link lengths depend on current traffic conditions.
We note that the algorithms of this paper are far from optimal since they
are single path algorithms in the sense that at any given time there is only
one path per origin-destination pair along which messages can travel. Better
performance can be achieved by allowing multiple paths as for example in the
optimization algorithm of Gallager 19] or its second derivative versions [2].
On the other hand the hardware limitations of some of the presently existing
networks including the ARPANET preclude the use of such more sophisticated
algorithms. Furthermore, we feel that the mere fact that the algorithm has
been successfully implemented in a network as interesting and influential
as the ARPANET makes it worthy of analysis and investigation. This is
reinforced by the fact that the behavior exhibited by the algorithm is
quite interesting and can ,se nontrivial design problems.
The paper is organized as follows:
In Section 2 we provide a deterministic finite state Markov chain
framework for studying a simple version of the algorithm. We show that for
ring networks the algorithm may tend to oscillate between poor routing paths
and become itself a major contributor to congestion. We also demonstrate
how the use of a bias factor can provide a mechanism for damping oscillations
as confirmed by experience with the original ARPANET algorithm.
4
The finite state model does not lend itself to analysis of more sophisticated
routing schemes and more general network topologies. We consequently introauce
in Section 3 a model of a ring network with a continuum of nodes and a single
destination. This allows us to employ techniques of stability analysis of
discrete-time systems with continuous state space, and enables us to further
quantify the relationship between choice of link lengths and algorithmic
behavior.
The analysis of Section 3 focuses primarily on the effect of using a
bias factor as a damping mechanism. In Section 4 we show that oscillations
can also be damped effectively by making the link lengths dependent on
several preceding routing paths via some averaging mechanism such as a fading
memory scheme or asynchronous link length updating. To our knowledge the
fact that averaging can provide a damping mechanism in a shortest path algorithm
has not been noticed earlier and in fact when we originally approached this
problem at BBN, Inc. there was considerable concern regarding its effect on
algorithmic behavior. It is now believed that the significant degree of
averaging inherently present in the SPF algorithm is in large measure responsible
for the stable dynamic behavior observed in experiments conducted thus far
(ll].
The analysis of Sections 2-4 focuses on ring networks. The ring
topology is central for the extension of our earlier results to more complex
network topologies. This extension is carried out in Section 5 under the
assumption that an equilibrium routing exists. However, by contrast with
ring networks, an equilibrium routing need not always exist for more complex
topologies. We demonstrate via example the mechanism by which such a
phenomenon can occur.
The results and analysis of the present paper can be generalized to
the case where there are more than one destinations. This analysis is
5
straightforward but considerably more complex technically and may be found in (3].
The continuous node model of Sections 3-5 may be criticized on the grounds
that it is unrealistic. On the other hand it is very difficult to provide an
extensive analysis of a more realistic finite node network model. In particular,
it appears impossible to demonstrate the effect of averaging in such a context.
Furthermore we believe that the realism of any algorithmic model must be judged
on the basis of the validity of the conclusions it provides regarding the
behavior of the related practical algorithm. These conclusions in our case
have been verified by extensive numerical experiments with finite node net-
works [4], [3]. In particular the validity of our qualitative results
regarding the role of a bias factor and averaging as damping mechanisms
have been amply demonstrated.
2. A Finite State Markov Chain Model
Consider a communication network with nodes denoted by 1,2,..., N and
directed links denoted by (i,Z) where i is the head node and t is the tail
node. We consider the following algorithm for periodically updating paths
for routing messages.
(A) At the beginning of every time period a nonegative length Dit of
every link (i,k) becomes availabe to each node. Based on these lengths each
node computes a shortest path to each destination and routes messages over
that path during the period.
The standing assumption for algorithm (A) is that the lengths Dit used
in computation of a new shortest path depend exclusively on one or more
preceding shortest paths. This dependence is deterministic via a rule that
for the moment we leave unspecified. As an example D may represent some
6
measure of average delay per message on link (i,k) during one or more preceding
periods perhaps with an added bias factor -a scheme currently implemented in
the ARPANET [1], [11]. By assuming that the dependence of Di on previous
shortest paths is deterministic we also implicitly assume that the input traffic
originating at each node is a stationary stochastic process whose ensemble para-
meters can be adequately measured by time averages. This assumption is not
valid, of course, in practice but is a reasonable approximation to the situation
where the time constant of traffic statistic variations is large relative to
the shortest path updating period (a quasistatic assumption, cf. [9]).
Consider first algorithm (A) applied to a given network for the case where
the lengths Dit depend exclusively on the preceding shortest path. Assume
also that the shortest path algorithm has a fixed rule for breaking ties
between equidistant paths. Then each shortest path uniquely determines the
next shortest path. There is a finite number of possible shortest paths
(also referred to as routings) which we denote by R, R2 ...... RM where M
is some integer. To any initial routing say Ri , there corresponds a1
unique sequence of subsequent routings R. , Ri.... Thus eventually some
routing will be repeated (say R. = R. ), and once this happens the routingIk k+n
sequence will become periodic. Thus starting at R. the algorithm will0
eventually end up cycling through Ri ,...R. k+nl. Of course it is possible
that Ri itself is part of the cycle (k=O), and that the cycle consists of0
a single routing (n=l) in which case the algorithm stabilizes at that
routing.
The model just described is one of a deterministric finite state Markov
chain with states Rl,...RM. From Markov chain theory or by elementary
reasoning it follows that the set of all routings {R ,... R M can be parti-
1 14
7
tioned into a collection of cycles (or ergodic classes), and a collection of
transient routings. If the initial routing is transient it is never repeated
by the algorithm, and if it is part of a cycle the algorithm returns to it
periodically. More than one cycles may exist. Furthermore, each transient
routing leads to a unique cycle.
When the lengths Dij depend on a fixed number (say m) of preceding routings,
a finite state model for the algorithm can be similarly constructed whereby
the state space of the model is the set of all m-tuples of routings. Similarly
the state space can be partitioned into cycles and transient states. Analysis
of such a model is naturally more difficult in view of the increased size
of the state space, and this is more so if D. depends on all precedinS
routings in which case a countable state Markov chain model is necessary.
In what follows in this section we will restrict attention to the case
of a ring network with N nodes shown in Figure 1. Node N is the only
destination and all links are bidirectional. By reversing the directions
of flow and the role of origins and destinations the subsequent model can
be converted to one with a single origin and many destinations. The
traffic input originating
N rN-2
N-2
ri-1 ri ri 1
8
at node i and destined for N is denoted by r.. The routing R. , i = 1,...,Ni, ~ 1
is the one for which all nodes j < i route their traffic in the clockwise
direction and all nodes j > i route their traffic in the counterclockwise
direction
0
ROUTING R1
Figure 2
as shown in Figure 2. Given a routing Ri, the flows on each undirected
link (j-l,j) in the clockwise and counterclockwise direction are denoted
by f (i) and f+(i) respectively and are given by
f~i)i .I 0 if i a j
ri + ri 2 + ... +r if j < i
- ri+r i +1 + ' ' ' + rj 'l if i ,< j i
0 if isi.
We will consider the case where the length D i of a link (i,j) is given
by an equation of the form
(1) Dit = d(fix)
where fi£ is the flow on link(ij) during the preceding period and
d is a real valued, continuously differentiable and monotonically
increasing function of flow with d(O) > 0. For simplicity we assume that
the function d is the same for all links but this does not affect materially
the analysis that follows. Since the flow fit depends only on the preceding
routing the same is true for the length Di . It appears that this simplest
of all possible situations is the only one that can be analyzed effectively
in a finite node network context. The practical situation where D is
taken to be the average time delay for a message to traverse link (i,i) can
be reasonably modelled by a function d of the form
'1(2) d(f P +T + (f
where
P. = Average processing plus propagation delay per messageit
T i = Average transmission delay per message
Qix (f i)= Average queuing delay per message when the average
flow on link (i,k) is f
The quantities P and T are independent of the flow f while the
dependence of Q it on f is determined by the statistics of the traffic
arriving at i and routed through Z. If these statistics can be adequately
modelled by an M/M/l queue then QiZ takes the form [6], [7)
f(f(3) Qi (fi£i i t c i t f i k
10
where C is the transmission capacity of link (i,k). We mention, however,
that on the basis of experiments conducted thus far it is unclear whether
the average delay per message in the ARPANET can indeed by modelled as in
(2). This may be due to peculiarities of the ARPANET hardware which are
little understood at present. We now define the shortest path algorithm
Given a routing Ri we define the distances D (i,j), D +(i,j) of node j to
the destination in the counterclockwise and clockwise directions respectively
by
D (i,j) = d[f9(i)]k=i
I ND (i,j) = dtf(i)J.
k =j+l k
+ iIf D-(i,i) = D (i,i) then the algorithm sets the next routing to R.. If
D- (i,i) / D+ (i,i) the algorithm sets the next routing to R where the node nn
is such that
D-(i,j) > D+(i,j) for j > n
D (i,j) < D (ij) for j < n
It can be easily shown that if D-(i,i) D+ (i,i) the next routing R isn
uniquely determined by the relations above. Given an initial routing R0
we consider the sequence of successive routings Rl 2 k .kl ,
generated by the algorithm.
The quantity d(0) may be viewed as a bias factor. It represents link
length at zero flow. The following proposition shows that if d(0) = 0 and
• 0 R1the first two routings are different, i.e. R 0 R then the algorithm ends
up oscillating between the two extreme routings R1 and RN which is the worst
possible behavior that can occur. In the context of (2) the case d(0) = 0
corresponds to the situation where the processing and transmission delays
Pit and Tit are negligible relative to the queuing delay Qit.
rr11
0 1Proposition 1: Let d(O) =0 and assume that R 0R Then there exists an
index k such that for all k-ek either Rk R and Rk+l =RN k =Rand
k+lR = R1 .
Proof: Let Ri be a routing and assume that the routing subsequent to Ri is
R with n # i. For concreteness assume that n < i. We will show that eithern
i = N or else the routing subsequent to R is R. with j i.
If i #N then since Rn is the routing subsequent to Ri we have
(4) D-(i,n-l) < D+(i,n-l) = D +(i,i).
We also have
(5) D (i,i) D+(n,i),
(6) D (n,i) D; D (i,n-l).
From (4) - (6) we have
D+(n,i) e D(i,i) -; D-(i,n-l) > D-(n,i)
so finally
D+(n,i) > D-(n,i).
It follows that in the routing Rj which is subsequent to Rn, node i will
switch his traffic to the clockwise direction so that j > i.
We can show using a very similar argument that if n > i then
either iwl or else the routing subsequent to Rn is R with j < i.
Thus we have that the number of nodes that lie between two
successive routings is increasing at each iteration if none of these
routings is R or RN . On the other hand if the current routing is R1 or
RN then the next routing will clearly be RN or RI respectively. This proves
12
the proposition. Q.E.D.
0 1Notice that, if d(O) 0, the situation R -R can only occur if
D'(ii) -D +(ii) where i is the node for which R0 =R i Thus if we add any
> 0 to any one of the node inputs we will have R # R 1 and the algorithm
will again end up oscillating between R1 and RN. We provide an example
illustrating the result of Proposition 1. Several additional examples
involving more general topologies and multiple destinations may be found
in [41.
Example: Consider a 16-node ring network where node 16 is the destination.
Letr. = 1 for i = 1,... ,7,9,...,15 and r8 = C > 0. If £ = 0 and the initial
routing is R. then by symmetry all subsequent routings equal R.. If C
is very small but positive then for the case where
d(f) = f
the sequence of generated routings is R8 , R1 0 , R3 , R16 , Rl, R1 6 , RI,...
This fact can be verified via a straightforward calculation in Figure 3
which shows the flow patterns corresponding to successive routings.
We now turn our attention to various notions of equilibria and
stability. We say that Ri is an equilibrium routing if
D-(i,i-1) < D+(i,i-1), and D +(i,i) < D-(i,i).
It follows from this definition that R. is an equilibrium routing if and1
only if it repeats itself via the shortest path algorithm.
We say that a node i is an equilibrium node if
D-(i,i) < D+(i,i), and D +(i+l,i) < D-(i+l,i)
In words a node i is an equilibrium node if he switches his traffic in
both cases where the routing is Ri and R.
_ _ A
Rs R16 L
0A 7 0 + 1+
O 4 O/A 3 E
00 0~
1st Ruting4th Routing
2nd Routing 5hRuIn
RIO1+
5 5 Fiur 0 t otn
14
We say that an equilibrium routing R. is locally stable if routing1
R generates either Ri or R. through the algorithm, and routing R
generates either Ri or Ri+ 1 . We say that an equilibrium node i is locally
stable if routing Ri generates Ri+ 1 via the algorithm, and routing Ri_ 1
generates R.. The definition of local stability is based on the idea that
when the algorithm starts "close enough to equilibrium" it should not lead to
a "growing" oscillation. The following proposition complements Proposition 1
and suggests that the bias level d(O) should exceed a certain positive value in
order for an equilibrium routing or node to be locally stable.
Proposition 2: a) An equilibrium routing Ri is locally stable if
r -1 N-1 r. N-1 -d(O) > max { 2 m£ , m
where
M = max{d'(f)If(i-l) < f < f, (i)} for 2 = 1.... ,i-l
= max{d' (f f+- (i) < f < f+ (i-l)} for = il..... N-1
= max{d'(f)If (i) < f < fi(i+l)} for Z = 1....,-2
i2 + = max{d'(f)lf+(i+l) < f < f+ (i)} for £= i,...,N-1
where d' (f) denotes the first derivative of d at f.
b) An equilibrium node i locally stable if
r. N-1d (0)> - E
-2 91-1
where
M = max{d'(f)lf2.(i) < f < f- . i~l)1 for Z =
m - max d'(f)lf+(i+l) < f < f+(i)} for I - i+l,...,N-l.
15
The proof of Proposition 2 involves a straightforward but lengthy
argument and will be ommitted. It can be found in [4].
Proposition 2 implies that in order to ensure local stability the bias
d(O) should exceed a level that depends strongly on the traffic conditions.
This level is proportional to the input at or near the equilibrium and to a
global measure of the derivative d' along the ring. Thus it may be necessary
to choose a value of d(O) which is large relative to r and d' in order to
ensure stability for a broad range of input traffic conditions. This can be
accomplished by adding a large constant to d. On the other hand this would
introduce a tendency in the algorithm to generate routings close to the min-
hop routing (i.e. one that selects route- according to minimum number of links
to the destination). As a result the algorithm would tend to be insensitive
to congestion. This tradeoff will be reencountered in the next section
The point of view that has been adopted in this section is one whereby
the algorithm is viewed as a dynamic system with a finite number of states
(the finite collection of possible routings). Unfortunately the study of
dynamic behavior and stability properties of such systems is notoriously
difficult. To begin with there is no accepted definition of equilibrium, and
in fact we saw that in the ring network context there are two types of
"equilibria" that are of interest - equilibrium routings and equilibrium nodes.
Furthermore there are no established methodological tools that can be helpful
in a finite state system framework. As a result our progress has been limited
to the results just discussed. We are thus motivated to consider approximation
of the discrete system with a continuous system having a continuum of states.
For such systems there is an effective and well developed stability theory
that can be utilized for analysis. We take this approach in the following
two sections where we introduce a network with a continuum of nodes. Despite
16
the radical nature of thissep the analysis provides informative results
and clarifies the role of averaging the effects of several past routings as
a means of damping oscillatory behavior. The validity of our approach is supported
by the fact that qualitative conclusions drawn from the continuous node model
have been verified computationally in finite node models.
3. A Continuous Model of a Ring Network
We consider a continuum of nodes arranged in a ring and sending traffic
to a single destination as shown in Figure 4.
1/4 3/4
t/2
Figure 4
Points on the ring are identified with their distance t from the destination
in the counterclockwise direction, where t is normalized to take values in
the interval [0,11. Traffic can move on the ring in both directions.
For every t in (0,1] we denote by r(t) the input density at t. The
meaning of the function r is that for any subinterval [tl,t 2 of (0,11 the
total input traffic originating at nodes in it1,t21 is
17
t2r (t) dt
t1
We assume that r is continuous on [0,11 and r(t) > 0 for at least on te (0,i)
Note that a network with a finite number of nodes can be modelled by a function
r containing impulses and such a function can be approximated by a continuous
function consisting of narrow triangular pulses of finite height. We are interested
in routings specified by points y in [0,1], where the flow splits, i.e.points
larger than y send their flow counterclockwise (or in the positive direction) and
points smaller than y send their flow clockwise (or in the negative direction).
To a given function r and routing y, there corresponds at every point t a flow+
in the Positive direction f (y.t), and a flow in the negative f'(yt) given by
t r(T)dT if y< t
(7) f+(y,t) =
0 if t<y
_ 0 if yf t
(a) f (y,t) -
SY r(T)dT if t < yt "
In order to introduce an algorithm such as (A) in the framework of
the continuous model we consider a function d mapping flows into the non-
negative real numbers. The meaning of d is that given a routing y and any
point t, the distances D- and D+ from t to the destination in the negative
and positive direction are given by
t
(9) D(yt) - df(y, ") ] d0
(10) D +(y,t) -~ d[f +(y,T))dr.t
We will assume that d is a monotonically increasing function of f with every-
where continuous derivative. We further assume that d(O) > 0. As Proposition 1
shows, the case where d(O) = 0 is not interesting from a practical point of
view.
We consider the following algorithm for generating routing sequences
{yk}:
(Al) Given a routing yk' the next routing Yk+l is the solution of
the equation
(11) D (yk, Yk+l = D+ (yk yk+l)
It will be shown as part of Proposition 3 that equation (11) has a
unique solution for every ykE[O,l]. Note that since we have
D (yk,t) < D(Ykyk+l) = D +(ykyk+l )< D+(ykt) if t < yk+l
and- + D(k
D (yk,t) > D (yk'yk+l) = D (yk'Yk+l) > D+(ykt) if t > Yk+l
it follows that a routing yk+l determined from (11) is such that every point t
routes its flow in the positive or negative direction according as
D (yklt) > D (yk,t) or D(yklt) < D+ (ykt), i.e. according to minimum distance
to the destination.
We say that y* £[O,l] is an equilibrium if
(12) D"(y*,y*) = D+(y*,y*)
We first show some preliminary results relating to existence and
optimality properties of equilibria:
Proposition 3: There exists a unique equilibrium y*E (0,1). Furthermore
equation (11) has a unique solution yk+l for every yk"
19
Proof: Using (9) and (10) we have for all y and t
(13) oD = d[f (yt)], 6D -d[f(y,t)].at a t"
We have d(O) >0 and d is monotonically increasing,so 6D (y,t) > 0 and
OD+(Y't) < 0. Thus for fixed y, the function D-(y,') is continuous,
monotonically increasing and satisfies D (y,O)=0, while the function
D+(y,.) is continuous monotonically decreasing and satisfies D +(y,l) =0.
Hence the equation D-(y,t) =D +(y,t) has a unique solution in t lying within
(0,1). Denote by g(y) the solution corresponding to y. The function
g:[0,1]-[O,l] can be easily shown to be continuous and, by Brower's fixed
point theorem ([8], p. 161), g has a fixed point y . This y is an
equilibrium. If there exist two equilibria yl and Y2 with yl < y 2' then
since d(f)> 0 for all f> 0, we must have
D-(y*,y* ) < D (yl,Y2) <D-(y2,Y2) =D (y2,Y2)
*+ +( * *D (yY) <D (y2 ,y*) <D lY ) *D (yll
which is impossible. Hence the equilibrium is unique. Q.E.D.
Proposition 4: The equilibrium minimizes over all y C [0,1] the expression
II I+J(y) = I> 1)f(y,t) ]dt f p[f- (y,t) ]dt
where p is any function satisfying for all f
(14) ' (f) d (f)
and p' denotes the first derivative of p.
Proof: The first derivativw J'(y) of J is given by
1 ;if (y,t) +Jll, , -t) -j~
(15) J,(Y) - I f +(y,t)]-f-- - (it + I f (y,t)] Df (y,t) dt. ly:01
20
It can be seen from (7) and (8) that
af + (y.t) -r(y) if y < t i
(16) y 0 if t < y
f 0 if y <t
(17)
r (y) i f t<'y
Combining equations (14) - (17) we obtain
j' (y) ffir(y)[ - fld[f+(y,t)Jdt+ 'od[f-(Y't)]dt]
or equivalently
J'(Y)- r(y)[D(y,y) -D +(y,y)].
If y is an equilibrium it can be seendiat we have
D (yy) S D+(yy) if y 'y
D (y,y) _D+(yy) if y > y
Thus J'(y) <0 if y < y*, J'(y) > 0 if y* < y, and J'(y*) = 0.
It follows that y* minimizes J. Q.E.D.
Proposition 4 shows that one can minimize the integral of
average delay over the ring by choosing the function d to be marginal
delay and by guaranteeing that the algorithm converges to an equilibrium.
The needto use marginal delays as link lenqths in order to minimize total
average delay has been pointed out earlier in a different algorithmic
context [9]. The following discussion, however, casts doubt as to whether
the algorithm will converge to an equilibrium when the link lengths are
21
chosen to be the marginal delays. In any case Proposition 4 suggests that
convergence of the algorithm to an equilibrium is desirable since a function
p satisfying (14) is monotonically increasing and convex and hence an
equilibrium will at least be a reasonably good routing even if it is sub-
optimal in terms of a particular design objective.
we now consider the convergence properties of the algorithm. For
any yc[0,1] we denote by g(y) the unique solution in t of the equation
D-(y,t) = D +(y,t) (c.f.Proposition 1). Thus Algorithm (Al) can be written 4(18) Yk+l = g (Yk)
We have for all yE[0,1]
g(y)
(19) D [y,g(y)] d[f-(y,t)]dt d[f (y,t)ldt = D+[y,g(y)]
dgy)
We evaluate the first derivative g'(y) d y)for yS(0,1). Differentiationdy
in (19) yields
g(y)
f d~f (y,t)]dt +- d[f (y,g(y))]g'(y)0
or = g~y)d'(f (y,t)]dt - dif (y,g(y))]g' (y)
fd'[f+(yt)]dt -jq d'[f (y,t)]dt
(20) g' (y) Z ____________
d[f-(y,g(y))] + d[f+(y,g(y)))]
We have for t # y
(21) 3d(f+(y,t)] = d,+ff+(y,t)) f (y,t)
-(21) af- (y,t)
(22) ad[f (y,t)] = d'[f (y,t)] - y(
3y
22
combining (20) - (22) with (16), (17) we obtain
min{y,g (y) } - +(23) ( y) [ d[f-(y,t)]dt + 4max{yg(y} d'[f+(y,t)ldt]
d[f-(y,g(y))] + d[f+(y,g(y))]
+at the equilibrium y* we have y* = g(y*) and f (y*,y*) = f (y*,y*) = 0, so
(23) yields
r(y*)[ J'df(Y*,t)Idt + f d'ff+(y*,t)]dt(24) g' (y*) = -:
2d(0)
By using a theorem of Ostrowski ([8], pp. 300-301) we can state the following
local convergence and rate of convergence result for algorithm (Al).
Proposition 5: Let y* be the equilibrium. Then if Ig' (Y*)f < 1 or
equivalently
() d r(y*)[ dhf(y*,t)]dt +J*d'[f+(y*,t)]dt](25) d (0) > I
2
there exists an open interval I containing y* such that if y0 eI the sequence0P
{ykI generated by algorithm (Al) remains in I and converges to y*. Further-
more if yk 3 y * for all k there holds
k1
(26) lira sup _Yk+l I = lim sup ly - y* I = g'(Y*)l(26)y li T y217 kim k 1
When the equilibrium y* has the property specified in the first conclusion
of Proposition 5 we say that it is locally stable. If Ig' (y*) > 1 then the
linearized system corresponding to yk+l = g(yk ) is unstable, so the
algorithm tends to diverge from y* when started close to it. Notice the
similarity of equation (25) with the corresponding local stability conditions
for finite node networks (cf. Proposition 2).
23
A sufficient condition for global convergence of algorithm (Al)
can be obtained by requiring that g be a contraction mapping, i.e. for some
pE(O,I) there holds
(27) Ig() -y * l : y-y * I, , VyE[0,1].
From Taylor's theorem and the fact g' (y) < 0 we have
g(Y) - Y*= I '(z)dzl
Let
j=max d' (f)
0O-r f j r (t) dt
0
From (23) we obtain for all z
Ig, (Z) z ~~zf - g (Z)2d(0)
Thus (27) is satisfied if
y__y <I
sup 2d(.) y-yyE[0,l 2
y y
or equivalently ifI r(z) 1 Iz -g(,Z)[dz
(28) d(0)> sup *y.
yE[O,l] y-yy y
24
This will be true in particular if
(29) d(0) > OR2
where R=max r(t). The conclusions of the preceding discussion are0 ' t : i
summarized in the following proposition.
Proposition 6: If condition (28) or the stronger condition (29) holds,
every sequence {y I generated by algorithm (Al) converges to the equilibrium y*.
When the equilibrium y* has the property specified in Proposition 6 we
say that it is globally stable.
In order to put the results obtained thus far in better perspective
let us write d(f) as
d(f) = a + d(f)
where a = d(0) represents the bias factor. For fixed input density r we
have that to each positive value of bias a there corresponds an equilibrium
y The equilibrium is locally stable for a satisfying [cf.(25)]
(30) L > X
2
and globally stable fora satisfying [cf.(29)]
BR
(31) a > R
As 0 increases the corresponding equilibria tend to become stable. Further-
more from (24) and (2?) it can be seen that the speed of convergence of the
algorithm is accelerated as a increases. On the other hand it is easy to see
* 1that y- * as a w , ich in the context of the routing problem means that
L d
25
the algorithm becomes incr4 -singIy ir!vensitive to congestion as 00"o
Since in a practical situation we are interested in the stability Properties
of the algorithm for a broad range of inputs let us consider input densities
of the form
(32) r (t) = Ar(t)
where A is a positive parameter. Then it is clear that as X increases a larger
value of bias is necessary in order to stabilize the algorithm.
For example if d is of the form
(33) d(f) = ' + fn
where 3>0, n>O then from (3) and (31) we see that if r is changed to Ar
nas in (32), then the stabiliLv., threshold level of the bias is multiplied by X
Thus for fixed a and r there is a choice of X for which the corresponding
equilibrium is unstable. Incidentally the expression (33) for d has an
interesting property, namely,that the set of all possible equlibria {yaU>0}
as well as the set of all locally or globally stable equilibria is independent
of the level of input A and depends only on r. This is straightforward to
verify using (33) and the fact that if r is changed to Ar and a is changed
to Ana then the routing sequences generated by the algorithm are unaffected.
Choosing the Bias as a Function of the Current Routing
Since stability of the algorithm depends strongly on the level of bias
and the level of input we are motivated to consider schemes where the bias
is not held fixed but is rather adjusted adaptively on the basis of currently
available information. An interesting scheme is to use a length function
of the form
26
if (Y)+ d(t)
wICJ d 1~i on , i iuousiy 4ifferentiable, monotonically increasing function
with ik0) C, .i-l x(y) i.. taken to be some monotonically nondecreasing
, . f , )ldt + d f (y,t)ldt.
"C'" '--n C> 2 untior. of the form
.,-y) + Y,[D1T(Y)]2
L i,:0 X.:Iiimontally determined nonnegative constants
;or~~~ms1 ku a,<. ni ,ontext of a finite node network with not necessarily
r srotc~ui , sche like this can be very easily implemented. In this
( c . ho caiciubat,. as the sum of all reported link "delays"
S 'iis (y) can ,e computed by each node via a formula such as
(34) and t ,o link lenqt can he computed as Di= (y) + d(fi)-
,; ih , e thc !vi'vp, just described can be analyzed along similar lines
a5 ', ri ir ii. t : 2.;t 2 L. it hac be-en tested in quite extensive numerical
t:xpti>lwInt:. ir,.IJie I finlt. node networks and it was shown to have very
fac~ >o y pcrr -manhc. [4] , [3]. This can be attributed to the fact that
, I. [vcif bia'; incre s or decreases with the level of input thus
'J~i n olitomine ,alinq with respect to input level. In fact it can be
!hia i'Lar it d ha.; the. form d(f) = Bfn where >O, n>O and we choose
,4 y) Y1D,,(y) wherc Y( i... for every input density function of the form
\c t), 4, th. :,'. onerit-i by the algorithm do not depend on X.
If
27
4. Averaging the Effect of Several Routings
In this section we show that the stability properties of the shortest
path algorithm of the preceding section can be improved if link lengths suitably
depend on flows corresponding to several past routings. There are several
possibilities along these lines. Some examples are as follows:
a) Averaging over the present and the past n routings.
Given a sequence of past routings ykYk-l,..., we define for any
t in [0,1] "averaged" distances to 0 and 1 by
0 i=0
(36) +(kyy k-n't n+l i=0 -i
1 n
Thus distances are calculated by integrating -4j i 0 d[f(yk i, T )Iwhich is
an averaged length over the routings yk' ..... k-n' in place of d(f(ykT)]
which is the length corresponding to the last routing.
The new routing yk+ I is obtained from the equation
(3) -ykk_ .... ,k-n, Yk+l ) =D +(ykYk-l,... ,Yk-nYk+l "
It is easily seen that this defines uniquely k in terms of ykYkl,
As earlier we write the corresponding equation as
(8) Yk+I = g(Yk'Yk-I ..... Yk-n ) "
A routing y is said to be an equilibrium if
jL
28
y g(y ,y ,...,Y ).
It is clear that y is an equilibrium in this sense for a given bias level
if and only if it is an oquiii.brimn in the sense given in the preceding
section.
We can define local stability of y in the obvious way. We have
that y is locally stable ii it is also a stable equilibrium of equation
(38) linearized around y (5cc [8] p. 353). It is a known fact that this is
true if all roots of the characteristic polynomial
C P n+l 1 3-L n Og(v) n-iC (p ) p k p - P - ''
k k- Ik-n+l 6Yk-n
lie inside the unit circle, (i.e. have modulus less than unity). We cal-
culate the derivatives
We have for a>0 similarly as earlier for every i
* y1
2,1 ( --+ 'j d'[f-(y*,t)]dt +j d [f (y*,t)]dt}k-i 2d() n+ 0 y
Define
(39) d r' [ ' [f-(y*,t)]dt + J , d,[f +(y,t)]dt}2d(0) "0 y
Note that, from Proposition 5, y is locally stable for algorithm (Al) if
p<l. The characteristic polynomial can be written as
(40) () n+ I + n _ n- I + _'40 '~) + 1-I P i-I p " + n+l n+l
We now use ti,( iollowing fact:
Lemma: Let g be a po!itiv, scalar and n be a positive integer. The roots of
the polynomial
29
n+ n - n-Ip +.p +lp +...+p+;
lie inside the unit circle if and only if >".
Proof: This result can be shown by straightforward application of Jury's
stability test ([10], p. 97-98). Q.E.D.
We now apply the result of the lemma to our problem. We have
that the equilibrium y will be locally stable if
n + I.
It follows using (39) that in the averaged algorithm the bias level must
satisfy{* y 1
r(y If dt + d' f(,.*,t)] dt
d(0) > 0 y2 (n+l)
in order for the corresponding equilibrium y to be stable. if we compare
this with the earlier algorithm [ci . (2:,)] we sve that in the averaged
algorithm the bias threshold level for stability is r duced by the factor
I1 over the one of al,;orithn (Al). For a given traittic input, and anyn+ I
given bias level the corresponding equilibrium can be made stable by
averaging delays over a su'ficiently larpe number (A periods.
Regarding rate ol convergence, 0 trowski's Theorem again applies.
We have from the proof of Th. 10.1.3 of [8] that given any >0 there
exists a norm on R :;uh that if y 11 y for all
(Y~k+1- Y, ... 'kn+lY )
lim sup k; ,*l P(n)4 fk .... yk-Y , .. -y )
where P(l4,n) is the rnayimi:w r,,t modulus oi the biaractcristic polynomial
C(p) of (40). It can be sUeen that for fixed n we have ¢ (i,n) 0 as - 0.
30
If are the roots of C(p) we have p . . =Pn- so thatP l nlth
P(a,n) t% q ) It follows that for fixed L we have P(4,n)-l as n- - ,
so that the rate of convergence deteriorates as n-'. Thus too much damp-
ing can slow down the speed of convergence of thu algorithm.
b) Fading Memory SchemeThis scheme is similar to the preceding one except that the lengths
corresponding to all past routings are averaged via a fading memory scheme. Given
the sequence of all past routings yk'lyk l ... }, the next routing Yk+l is
determined as the solution of the equation
i~ +
(41) j 6( t)dt f C (t)dt
0 Yk+lk
where 6k and 6+ are obtained by the following recursive fading memoryk k
scheme with decay factor E[OI)
6 k(t) = 06k (t) + (I - )d [f-(ykt)J
+ (t) + (1 - )d~f( 0t)]k~t k 6kI [f(k) "
Alternatively we can write
k(42) 6k(t) = (- ) i Dk-id[f-(yilt)]
kk
Let us write the solution of (46) as
(44) Yk+l m g(Yk'Yk-l1 . ... )"
Let us also consider the linear system obtained by formail linearization of*
(44) around the equilibrium y . We have similarly as earlier that this
linearized system is
2S(45) Yk+l f- ( )y + Yk[+ k- + "'']"
31
where p is given by (39). Let us denote
2Zk+l Yk + PYk-I + r Yk-2 +
Then we have for all k
(46) Yk+l = - 40 -")Yk - k( - 0)'z k
(47) Zk+l = Yk + Zk
and it follows that the linearized system (45) is in effect the two-
dimensional system described by (46) and (47). This latter system is stable
if both eigenvalues of the system matrix
lie within the unit circle. These two eigenvalues can be calculated to be
0 and B- (l- ). It follows that the linearized system is stable if
Although we do not provide a proof, it is possible to establish rigorously
that stability of the linearized system (45) implies local stability of
the algorithm (44) and thus we have the result that the threshold value of
bias, for stability in the f: ling memory scheme is reduced by the factor
-L1 over the one of algorithm (Al). The optimal speed of convergence is
obtained when the eigenvalue - (l - 0) equals zero in which case a super-
linear rate of convergence is obtained. This is so when =--- For other
values of in the interval , ) the rate of convergence is linear, andS+
32
for a-< the equilibrium is unstable. As ( is increased from the optimal
value towards unity the rate of convergence deteriorates.l+ji
c) Asynchronous Length Reporting
This type of scheme is patterned after a shortest path routing algorithm
where nodes report asynchronously the lengths of their outgoing links and the
shortest paths are updated after each report. The set of nodes 10,1) is
partitioned into n subsets which we call SIS .... ,Sn At some time, say 0,
the nodes in S1 report their lengths averaged over the flows corresponding to
the preceding n routings and a routing update takes place. Then at time
a > 0 the nodes in S2 do the same thing. Similarly, for i = 1,..., n-l, at
time (01 + a + ... + 0.) the nodes in S do the same thing. At time1 2 ii+l
(a + 2 +...+O n ) the nodes in S again report their lengths, an updating1 2 n1
takes place and the process is repeated. This type of asynchronous operation
is currently in use in the ARPANET [4] where, in a finite node network
context, S. consists of a single node for all i. There are also otheri
variations of asynchronous operation involving for example averaging over
all preceding routings via a fading memory scheme. This type of algorithm
is described and tested computationally in [4] and [3]. The analysis of
all these schemes is very similar as that of the averaging schemes described
earlier in this section. The details are quite messy and may be found in
[4], where it is shown, via analysis and computational experiment, that
asynchronous operation has a .ubstantial beneficial effect on the stability
properties of the short(st path algorithm.
5. The Case of a Network with an Arbitrary Topology
The extension of the continuous model to the case of a network with
33
arbitrary topology is quite straiqhtforward. However, the notation required
for a precise mathematical description is very cumbersome and tends to cloud
the main ideas. For this reason our pr scntation will be somewhat informal.
Consider the case of an undirected network with a single destination.
Let r be the input density function mapping points on the undirectred links
of the network to the nonneqative real numbers. The mcaning of r again is
that, given any interval I on a link, the total traffic input oriqinatinq
at this interval is the inteqral of r over I. We view the set of points on
the network as a subset of a Euclidean space of appropriate dimension, and
assume that r is a continuous function. In order to consider notions of
length we associate with each undirected link (i,P) two directions i- and
k--i. (There may be more than one links connecting a pair of nodes within
our framework. When we refer to a link (i,9) we mean a particular link
connecting i and Z and specify further when there is danger of confusion).
A length function 6 is a function which assigns to each point on an undirected
link (ij) two nonnegative numbers one associated with the direction i*V
and the other associated with tlic2 direction Q i. We assume that is piece-
wise continuous along every link in each direction. The meaning of 6 i's
that given any two points on a link (i,9) their dki tance in the direction 1 *
is obtained by integrating S as defined in that direction between the two,
points. The distance in the opposite direction -i' is definod ana]oou. ]y.
Similarly we can consider path,; between Point; on possibly different links
and define their long th in one or the other irection.
We now associate to a ivon lenqtIh function a shortest path of t-Vi ay
point, and an as!;r-iat,, r' tit irni. W,* ,j!:!7jmf that , is everywher, ;o : ,i ,.
(;ivn any point w-, tin: dr t- i t ion e, pt ; to th' ,lortimat (, n n
34
their associated distances specified by the function 6. A path of minimum
distance is referred to as a shortest path from the point to the destination,
and the corresponding distance is referred to as the shortest distance of
the point to the destination. The routing corresponding to 6 is the set
of points for which there are more than one equidistant paths to the
destination. A routing is said to be regular if it does not contain any
nodes of the network, otherwise it is said to be singular.
Given the function 6 , a shortest path of each point and the corres-
ponding routing can be constructed in a simple manner along similar lines 4as for usual networks. We first construct a shortest path tree for the
network in the usual manner by using as(directed) link lengths those speci-
fled by the length function 6. (The length of the directed link (i,A) is
the Integral of 6 along (i,l) in the direction i-1). This gives us a
shortest path and the associated shortest distance for every point on the
shortest path tree including all the nodes of the network. A shortest path
for points on links that are not part of the shortest path tree can be
obtained as follows:
Let (i,A) be a link that is not on the tree. Let Di and DA be
the shortest distances of nodes i and A. The shortest distance of a point
t on (i,L) is
'iAD(t) -min ID i+ J 6 i()dt , D+ J 61 ()dt}
t t
where 6 is 6 in the direction A i and 6 is 6 in the direction 1 -.
It can be seen that the routing corresponding to 6 is regular if and only
if each (ordinary) node of the network has only one shortest path associated
with it. If a routing is regular then every one of its points Lies in the
t
35
"interior" of some link. Notice that the preceding construction shows that
a routing (regular or not) consists of (L - N + 1) points where L and N are
the number of undirected links and nodes respectively.
Given a shortest path tree and the corresponding routing constructed
as just described, we can define the flow corresponding to it. At each
point, say t, of a link (i,Z) there are two flows to consider (one of which
is zero); the flow in the direction i-* and the flow in the direction -i.
Each is defined in the natural way by integrating the input density function 4r over the portion of the network that lies "upstream" from the point t,
i.e. over the set of points the shortest paths of which meet t on their way
to the destination. At the points of a regular routing the flow is zero in
either direction. Notice that if 6 is such that the corresponding routing
is regular the flow is uniquely determined by 6 . Otherwise the flow will
depend not only on 6 but also on the shortest path tree selected.
Suppose we are given a monotonically increasing, continuously
differentiable function d mapping flow into the positive numbers. Given a
shortest path tree T corresponding to a length function 6 with routing Y we
can define a new length function & which assigns to points t in any one of
the two possible directions the length 6(t) = d[f(t)] where f(t) is the
flow at t corresponding to 6 and T in the appropriate direction. The
corresponding routing is denoted Y. Note that if Y is singular then X and
Y depend not only on S but also on T. If Y is regular then Y is uniquely
determined by (S
We are now in a po,ition to define an algorithm cimi lar to the one of
Section 3. Given a ],,ngth function 6 and a corresponding shortest pathn
tree T and routing Y W the next le'ngth function i ; 7 with Y0 1(-orl'e!i,' inI' Y 1 (
routing Y1 .0 A T:hor test path tre:e T I 001 *,j 'end ng to 1 5 iS 1 ], et (,
36
and is used to define similarly 2' Y 2 and T Similarly the algorithm
generates 6k, Yk and Tk for all k.
We say that a routing Y* corresponding to a length function 6* and
shortest path tree T* is an equilibrium routing if 6 = (" and Y = Y*.
Contrary to the case of a ring network where we were able to prove
existence of an equilibrium, in general there need not exist an equilibrium.
This fact is demonstrated in the following example and provides an indication
of the complexity of the dynamic phenomena that we are investigating.
Example: Consider the network shown in Figure 5. 4
i:
/ i
\\
Figure 5
There are two nodes 1 and 2 and th~ree links connecting them denoted by
A,B,C. Node 2 is the destination. Points on A,B, and C are parameterized
by their Euclidean distance Lo the destination. The Euclidean lengths of
A,B and C are all taken equal to unity. Let the input density function be
as follows
37
For link A: r(t) E 1, VtE[o,1]
For link B: r(t) m rB) VtE[o,i]
For link C: r(t) E rc, VtE[0,1].
We assume that lr B : r l<r Let
d(f) = C' + f
where Y > 0 is the bias factor. K
In view of the fact 1 < rB < rc, l<rc, it is clear that an equilibrium
routing cannot contain a point in the interior of link A, while it must
contain a point in the interior of link C. We consider two cases:
Case 1: r B = 1. Then an equilibrium routing cannot contain a point in the
interior of link B so the only candidate for equilibrium are the two types
of singular routings shown in Figure 6. In routings Y and Y the incoming
traffic at node 1 is routed through link A and link B respectively. None
of the two routings can be an equilibrium. In routing Y1 there will be points
in the interior of link A which will have a shorter distance to the destination
(corresponding to Y1 ) through link B rather than through A, and the reverse
situation occures in routing Y Notice that this argument makes use only
of the magnitude of r and r and is independent of the form of the function d.B C
Case 2: l<r Then it can be seen that the only candidates for equilibria
are routings of the form shown in Figure 7. Each equilibrium routing candidate
38
is specified by the points yB' yce[Ol] where the flow separates on links B and
C. We have that the distances D+(yB), D-(yB) of Y b corresponding to routing
(yBYc) along the counterclockwise and clockwise paths respectively are given
by
D (y B B+rB J 0 (yB-t)dt0
+J [r B(I - Y + r C(1 - + (1 - t)Id t
If (yBPyC) is an equilibrium we must have
D"(yB) = D+(yB)
which after some calculation can be written as
(48) 2(O+rB)(1-yB) +rC( - yC) =B 2
By symmetry the equation D-(y ) =D+(y c) can be written as
rC - 1
(49) rB(l - yB) + 2 ('+ rc) ( -y) C
Equations (48) and (49) are in fact necessary and sufficient conditions for
(yB yC) to be an equilibrium routing. Thus there exists an equilibrium* *:
routing if and only if the solution (yB3yC) to these equations satisfies
yBE[0,11, y*E[0,11. After some calculation, this condition can be shown to
be equivalent to
rC( 2rB - r C I)(50) B - 1)
39
3.i
A B ( A B
Routing YI Routing Y2
Figure E
F//
Figure 7
40
If 2rB > rC + 1 then for every level of bias there exists an equilibrium rout-
ing (yB' YC )" If however 2 rB < rc + 1 then there exists an equilibrium only
for a above the threshold level indicated in (50).
The preceding example shows that existence of an equilibrium can depend
on both the level of bias and the input density function. Furthermore, it
may happen that, for a given input density function, no value of bias can be
found for which an equilibrium exists. This last phenomenon is of a singular
nature and is due to the fact that the Euclidean lengths of links A,B, and C
are all equal to unity. To see this consider the routing Y, corresponding
to the length function 6+ (t) H 1, 6-(t) E 1. The routing Y is analogous
to the min-hop routing in discrete node networks, and can be associated with
infinite level of bias. It is an equilibrium routing for the case d(f) £ 1.
If Y is a regular routing, i.e. each node has a unique minimum Euclidean
distance path to the destination, then it is clear that, for any given input
function r, there exists a threshold level of bias a such that for all
a >cta regular equilibrium routing exists.
Characterizing the dynamic behavior of the algorithm in the absence
of an equilibrium is certainly an interesting problem but we have been un-
able to make much progress in this direction. Computational results for
finite node networks given in [31 suggest that the stability properties of
the algorithm are improved by high level of bias and averaging similarly
as in the presence of an equilibrium. In what follows in this section we
restrict attention to the case where a regular equilibrium routing exists.
Given a regular equilibrium routing Y* = ... ,y* considern
for j = 1,2,.. .,n the link (ij ,£) containing y* and the two shortest paths
from y* to the destination. A simple but fundamental observation is that
3
these two paths join at some point thiereby forming a ring of the type con-
41
sidered in Section 3. The zero point on this ring is the point where the two
paths join. Let e. be the Euclidean length of the ring containing y. For
j = 1,2,..., n we parameterize points on the ring containing yj by the number in
[O,e.] going from smaller to larger numbers as we traverse the ring in a chosen
direction similarly as in the previous two sections. Thus points y. on
the link (ij,.j) can and will be identified by the number in [O,e.] specifying*
their position on the ring corresponding to y. It is easy to see now that
given Y , any collection Y = {yl1Y2, .... yn} such that y. lies in the interior
of (ijZj) specifies a flow f through each point in the network that follows
the (ordinary) shortest path tree corresponding to 6 and Y and separates on
each link (ij.,.) in the two opposite directions at the point yj. This flow
defines a length function 6 via the relation 6 y(t) = d[f y(t)] in the direction
of the flow, and 6 yields in the manner described earlier a shortest path
tree and a routing denoted by g(Y). It is easy to show (using the regularity
of Y ) that if Y is sufficiently close to Y then the (ordinary) shortest path
tree corresponding to 6 is the same as the one corresponding to Y and that
the elements of the routing g(Y) lie on the links (ijZJ ).
The algorithm described earlier can now be redefined a
(51) Yk+l = g(Yk)"
The definition is local within a sufficiently small neighborhood of Y and
is associated with the (ordinary) shortest path tree corresponding to Y
and the associated parameterization of the ring subnetworks containing the
links (ij,9j ).
Similarly as in the preceding section we say that an equilibrium
42
y is locally stable if there is a neighborhood of Y (defined in terms of
the parameterization of the rings associated with y as discussed earlier),
such that the sequence [g(ykuj generated by (50) is well defined and con-
verges to Y for every choice of Y0 within this neighborhood.
In order for Y to be locally stable it is sufficient that the
nxn matrix be defined and have all its eigenvalues within the unit
circle. The computation of ai(Y is straightforwqrd along the lines• 21Y
of Section 3. We first introduce some notation. For J -1,2,...,n
let + denote the set of points tE[yj,ej] on the jth ring, and R
denote the set of points tE[O,yj] on the same ring. NJote that for every
J,m-1,...,n the direction of flow on R and R+ (or R em) mustyj ,ej Ymaem ymm
coinside if these sets have intersection with positive Lebesgue measure.
This implies that at least one of the sets R+ nR+ and R+ n R"YJ,e I Ymem Yj yme m
is either empty or has Lebesgue measure zero. Similarly at least one of
the sets R e nRy and Ry fRr+ is either empty or has LAbesguey ,e y ,me m y ,ej ym,e
measure zero. The equations defining g(Y) can be written as
Pf AIF' d, = . rf+(Y,t)]dt, J -
tJ ) +- - * J "" -t. . .
Rgj (Y),e 1 (Y),e
By differentiation with respect to ym we obtain similarly as earlier at the
equilibrium Y
(g (Y , r(y*)
(51)o 2d (0) Jm'
where
43
d'[f+.(Y ,t)ldt(521 eJm N R+* +
yj,e. Ym,em
"4-. N- d'[fj(Y*,t)]dty ej Ym m
+R +d' [f-(Y ,t)ldt
R n R-*
+ * . .
y.,e.i yM ,e
d'[f.(Y*,t)ldt
and f+(Y ,t), f.(Y ,t) are the flows on the jth ring in the positive and
negative directions. In view of the preceding discussion, at least two of
the integrals in (52) are zero for every j and m.
Let R be the diagonal matrix having r(y.) as jth diagonal element,
and let @ be the nxn matrix having as elements the scalars 8 .m* Then we have
ag(y 1 8R.(Y 2d(O)
We can show that the matrix 0 is negative semidefinite. Indeed the matrix -Q
is the Gram matrix associated with the functions
XR +* e t d'.F(Y ,) - XR-* (t) d'!f.(Y ,t) j =n,
y] i yj,e.
where Xs is the characteristic function of a set S (X (t) = 1 if ttS, X(t) = 0
otherwise). By using the fact that R is diagonal it can be shown that the eiqpnvalu
X 'A of -g(Y )are real and nonpositive. Consider the spectral radius
n ay.
44
= max . . kThen the equilibrium Y is locally stable for
(53) I < 1,
and hence there exists a threshold level for d(O) above which the corrrsp;:,uing
equilibrium is stable. Similarly as in the preceding section, we can show
that if a fading memory scheme with decay factor is used to average the
effects of past routings the equilibrium Y is locally stable if
(54) 0 < 1+_ _
and there is a value of which optimizes the rate of convergence. It is
also possible to show that the other forms of averaging the effects of
several past routings improve the stability properties of the algorithm.
For the purpose of aiding the reader in understanding the method of
calculation of the matrix 9g(Y) we provide an example.3Y
Example: Consider the network shown in Figure 8 where node 4 is the
destination, and assume that the regular routing {y1 , Y2, Y 3 j shown in
an equilibrium. The figure shows also the chosen positive direction on
*the ring corresponding to each Y
f
45
+ +
4
Figure 8
We calculate the symmetric matrix 9 with elements e. given by (52).
The interval between any two nodes i and A is denoted [i,L]. The interval
between some y and a node I is denoted [yi,£]. We have
---- d'[f(Y ,t)] d--J* d'[f(Y ,t) dt
[Y2l1] U [1,3] U [3,41 [Yl ' 4 1
2 - 2 U d'[f + (Y*t)I dt d'lf2(Y*,t) dt22 [ 2,
2] [23 ( 2,l U[13
33 - ,d'If3(Y ,t] dt-J'* d'[f 3 (Y ,t)] dt
[Y3,2] U [2,3] U [3,4] [Y33 4J
1 - j d'[fC(Y*,t) dt
023'[+ *
(2,31
Li 46013 = - dfl(Y*,t)] dt
[3,4]
Acknowledgement
This investigation was conducted while the author participated in
a design study of the new ARPANET routing algorithm at Bolt, Beranuk, rl.,
Newman, Inc., Cambridge, Massachusetts. This study was supported by ARPA
under Contract MDA 903-78-C-0129. The collaboration and discussions with
John McQuillan, Ira Richer, and Eric Rosen had a substantial influence on
the ideas of this paper. The research was completed at the University
of Illinois, Urbana with support from Grant NSF-ENG 77-15949, and at the
Massachusetts Institute of Technology with support from Grant ONR-N00014-75-c-1183.
4-7
1] J. M. MCQU I I'viO 1. P 1~ 11, ".. 1 ':1<7 I.I,. I Al tjoAbju thmImprovemnits. F ir :t 1.'I~ l . I'l I)D l r tt No. i8.) 3prepareI to! A'PA iii Ilk I
1 D.P. Tirtijik.,,, "Al :, 1 i~ t !11., I'? i '' ! In Networks",('oordin~itrd -(i- !,. . 1i:- Ii: ilii ;, Jun( i9j?8 1*A. 1nii;, I i'I~ingi tr~ >79 pp. 210-224.
[3] D.1'. Lirt ;'k r int~ 1'~t. 0 p I t- I I A 1,;(,t t 1111;
for ('ovnuicit iI, i -t' w(tk wit m1 it *'!- t:i it] oli'' 1i 0 / IEEECon!rin [ii 2 1 -1 ii, 1 p. q 1 .'7 7-13 3.
[41 .T.M. ii )l ti ,I 11 1 1 B t k , "'A], 1ANYTJ Pout icg
foi ARP-A ai, P,7,. ti> ' .
r ''fori~ani j' LT.:' lY' Ai I a Y tl hcr 3 , I F1 1P ib J; 1)1
I'rilntiro, hOil', ~l 1 i r!I I ff N .J. I
171 l u, ,i V-n . w m ini -W ! i i ' . ' - ' l 9; t. -
d. . 1t4. ltim q1
191 R. GIwi T III..... , l in .1 ut <i 1. '1 z' 24 T: Ii in;I A I iritm fo
0i -)I t; F''' Ju fI Ii:......... f. ho, J .v ,i ki v 120
N i. Y
11 1 ,!O ii~j Al(;)rit m fo
t . AP ' t;
: Proc. 1979 IEEE Cljnof!rt li e (,;i -'i' 1-1r -,.1 o, I 0
.; Ft. Lauderda, e, Fla Li , - i'", 9! -1,
WA6 9:45
D OY JNZc A F_', wL O"ITKItS Mi.
COK-tJN1FA- N LT ie .r
Department of flectfilal i r-,. -' .v,: . r s ter Zcienc
LaUorator-j I..,- :.!n - ar 2 S,, tsn
Abstract
This paper provides analysis ard computationa! . d--ia ti.e inate val P betaren t, and t in the
results relating to the dir.azsic behavior of shotte.t tlt 2
path routiny algorithms for store and forward co:"uri- cLzckw-se direction is def :!ed to be the setcation networks. A companion paper (1) focuses onnetworks with a single destination he ee wokt t t2 ,l if tl<t2
considers networks with multiple destindtivnf. - -
1. Introduction (t2,t 1 if ti>t2
This paper is a sequel to a companion paper Ill "which examines the dynamic behavior of shortest path for tI t2 we define Ptlt - t }-(t
routing algorithms for counication networks with a 2 1 2
single destination. We examine here the case where Notice tat we have for all tl.t 2 [O(l)
there are several destinations. Generally speaking, +the results obtained for single destir-tion network Rt t 2 2,
have extensions to the multiple destination case 1'2
although the analysis is considerably more complex.
Similarly as for single destination algocithus -' For each I - 1. N we a-- given a continuous function
find that the addition of a bias factor and the ri [0.I J 0,) .uch that r (0) - r1 ). We refer to
presence of an averaging mechanism have a stabilizin r, i-. N, a; the - : ,ensit-- function foreffect on dynamic behavior. Our rsults also suggest d I. Th nan; of ri is that for any two
that the presence of multiple destinations has a ... . ..da=ping effect on the dynamic behavior of the distinct points ti,t 2 an the ring the total traffic
algorithm. Similarly as in (1) we consider a network input 'Hat originates in R -* (Pt- ) a destinedoodel with a continuum of nodes and rely on methods tilt2 tiltof convergence analysis for discrete-time systems for x Iswith Euclidean state space. The possibilities for
extension of these results to finite node networks (eem limited. However, it appears that conclusons J rft)dt (f %(t)d)
on algorithmic behavior drawn fron analysis of tlhe t t2
continuous model remain at least qualitatively validfor finite node networks. The conjecture is supported w !cv Irtroeuc- n,tio.-.a of lenth ani distanceby results of ccmaptational experiments provided in al,,ng the rtrci. These notione will be considered inSection 3. Throughout the paper we ass une that the bth. t' c'Cikwiso and cointerclo,:kwise directions
&eader is famillia- with (1). which are also the negative and positive directions on[0.11. By a len-_,h fuactio:, we mean a function which
2. A Continuous Model of a Ring Network with assigns to each ;soint teO,lI two positive numfers 6*(t)
Multiple Destinationand (t), th- first associated with the positive
We consider a continuum of nodes arraned in a direction and tL- seond associated with the negative
ring and sending traffic to N points on the ring directi'n. Ths a length function can be identified
referred to as the destinations. A point is selected with the correspoonding pair (6+,6- )
.on the ring and is referred to as the orijin. Every
other point on the ring is Identified with Its In what fllo-' attention will be restricted to
tuclidean distance t from the origin in the ounter- lenqth fctinnis 5 - (6',6-
) for which 6" and 6- are
clockwise direction, where t is normalizel r- take piceise ccntl--. us on 10.1). Given such a length
values in the interval (0,i. The destinations are function wci the disrance AIt t ) Of any twoidentified with thcir di'ntances ,.. . from 12
t?e origin. We assume that p'nts *~ I t] I t ir tht. positive direct onki
0<X < x I .
Given any two distinct points ti , t2 on the rinq, th* ( J 6(t dt
interval R* t between t anJ t 2 i the ca'rnter- t L'2
clockwise direction is defined to b,! th- ,e ' . rti t. i-.rtarca P6 -(t IIt 2 of t1 end t2 In the
itt t r.;~t eirrelrnn is defkred b
t 2. . ) . (t)d (4)
It'l~l U Nt| if 1tlt
2I II I I
Given a length function (6+, C), it is clear that destint;.on points x 1,.. x. For every fixed tjf +-t)
there exists for every I a point z elol] such that and ,N)
z p' xi and f and f are continuous on (0l1) . Furtherore it
aa. 6( s .sn that for all I ad ye(0',IN
46|lx)- &,*(zi~xi). (5)
The point zi is unique except in the singular case -riz ) if teR t1) z
where zi 0 and z .1 both satisfy (5). Define ay i
Yi f dt, I - l,. n. (6)
ie
(0 if toWaNote that y1 is the counterclockwise Euclidean 32-Y1t)*
distance between xI and a, along the ring. Clearly we y( (1)
have 0<y 1 <l for all I. The vector Y - ( ' 1Y ; 1() if teI- tI -
defined by (5) and (6) is referred to as the shortestpath routing generated by (6+, 6-). It is easy to where z. is defined in terms of .y via (7).verify that Y has the property
Sizmilarly as in ill, we are given a continuously
X,* yl < X2 + Y2 < .<
X + YN diffe:e-tiable° monotonically increasing scalarfuncticn d of flow such that d(f)>G for all f>O.
more generally by a routing we mean an element (For si:licity we exclude the possibility d(O) . 0.e eIt was stown In [1) that unstable algorithmic behavior
of 10,1N
(the Cartesian product of N copies of 10,11). results in the single destination case if d(0) - 0).Given a routing Y - (yi" " YN
) consider for each I Given a routing Y we define the length function
for which y y' 1 the point z on the ring given by (6*, a-) corresponding to Y by
Sx, + Y, if X, + Y, <1 1 +
{x:-:y7} I t) - dlf*(Y.t)]. Ite0.11 - (14)
(715x, + Y, 1 if Xi I Yi >
I yt ~ ' tl t[,],(5
Define for each i - 1,...,N and tel0,1) the flow
t+(Yt) at t in the positive direction for destination We denote by q(Y) - [H1 (Y) .. (X) the shortest
I by pat-h routing generated by (. are interested in
0 if y -l, or yi1~ the algorithm
a(Yt)nd t 1k+l g(y )
te 8) where Y is a given initial ioating.If + i d 9,,311ndtasilt i'iwe say tht Ye is an .equilibrium routing if
Similarly define the flow f (Y.t) at t in the negative y* ; g(Y).
direction for destination I by Since g(Y) belongs to (0,1) for every routing T.it follows that an equilibrium Y* must belong to
0 if y, pl and teC N.( 2~ , ~ lo0l) Usk have the following proposition regardingt
' existeneC, uniqueness. and optimality of an
r (Odl if y1 il and te, *equil&ZrtrL. The proof i quite lengthy and has besef (It) - z1,t i i releqtad to Appendix A.
Proposition 1: There exists a unique equilibriumif y - 1. routinv ¥*. Furthelmore V* minimizes over all
Ye(0,1)"
the expression
Define also the total flows in the positive and nega- I I
tive direction at telo,13 by J(1) /ptf (Y.t)idt +f pt"f(Y.t)]dtN + 0-a
f+(Yt) - E f (Y.t) (10) where p is any function satisfying for all ri-I
f (Yt) I I f (Yt) (11)1-1 Ws tave assumed earlier that d(O)'O. If this
ass:tion is replsed by d(O)>O then Proposition INotice that for fixed T - (yl,...,y") the functions can still .e shown to hold excewpt for the uniqueness4(,.) and f (V,*) can be discontinuous only at the part. txistence of an equilibrium can be show by
usin ,AAutant's fixed point theorem (12). p. Vi) in
128~
place of Brower's theorem in the proof of App-,, . Then we can write the Jacobian (17) evaluated at ye kkThe proof of the optimality property of an eq ilP 'ugiven in Appendix A is also valid under the agr -tion - (22)d(O) > 0. (2
We now evaluate the Jacobian matrix We now establish that the matrix e is negative ssm i -
definite. Po- any set S denoted by X the characters-
g () ag (Y) tic fu;,ctio o! S, i.e. X (t) - I if teS, and X (-)is
if tos. It cm be seen that -e is the sum of twopositive se=44elinite matrices. The first matrix 1L
athe Graz matrx of the functions
i }¥ ~~ ~~X 10 ... i- .. I- . 3f
agN(Y
) 39 N(M in L1O,l1 (.31, p. 56), and the second matrix is the
ay 1 . . . .. ... ... Gram matrix of the functionsS3d(f+ (Y*,t)I
For any routing Ye(O.l)" define 9.(Y), i .. N, by X + (t)
the equation (cf. (7)]. i ziIi
Since every Gra matrix is positive seilAfnte it
i+9(Y) if x + 9 (Y)il follows l.st: 9 is negative secidefinite. Now the
ii ( Y )
" elgenvalues of -d .YD) are the same as thext+ 1 ( )-l if x 1 +g (P> eigenvalues of Ah! a . Dl DlGR1/ID which is
a negative semidfinite matrix. It follows that all
The equation defining gi(Y) is cf. (3) - (5). (14),(I)|eqgenvalucs I 1 ... of a re real end nonposi-,,
d[f-(Y,t)ldt - d(f+(Yt)ldt, tive. It is als possible to show that q(y h
9 (),xx i set of eigenvectors that form a basis for R", I.e. itis diagona--zablo. Let P ht the spectral radius of
Differentiation of this equation with respect to yjY
yields similarly as In M1 the following formulzo
-It lrtT.N vl* 411. 1T.kt(T)1l is locally stable if there exists -n~ghborhood JV of
I ¥* such that if yoe) then the sequence (Yk} generated
(18) by the algorithm 116) remains in N and converges to Y
Using Osto"swki's theorem (121. p. 300-301) and thewhere z4 and 24 are given by (7). At an equilLbirum fact that ha hs real gnvaluesnd i dia .a l"
(y1 ... y) we have izable we :a.n prove the following y .--. ....
Propositiz: 2: Am equilibrium Y - i i
, [ wi ... 4(locally stable if
S' '* ,'. '*'. , [1 it <(24)
cli n.. ,i .,&..u1 where 1 is defined by (23). urtherm if ( k t
(19) and Yk # " for all k there exists a norm " '
where for all i such that 1
#5 if .,Y i In I.1?z 3 i : - k-
Let N. ar-I L) De Lh. ),aisvna r (Z I- j + ;(df)
(d(f-(Y'.,2)I d df(Y*'/.. z
r , ertle -i .ie y - dIO) represents bia. It follows from thedia ooael ements. Let al, fl be the .y10etrl Nx, N prec.'- - a-ly-ts that a suff cetly high level of
vmatrix having elements bis w1,i -- oie a locally stahle equilibrium ard an
ar.j-r : q.nilax to the one of III shows that forsf,, :'y high level of bias the algorithm will con-
0.1 ~ ~ ~ ~ ~ ~ e j.. ' H 1 ~: .: . t-~ an #7vilIirt5 for every Initial routing I.
Le,,,-.
When there is a single destination (N-I) we have It follows from the fact <0(and (23) thatY * , . : .*Y . : -t 0o l w ar od th e d e n m i n t o i n<_ ( 1 9 ) a2 )
f(Yzi) - f+(Yz) -0 and the denominator in (19) the linearized system is stable ifequals 2a. When there are more than one destinations 1 + Bwe will usually have f(Y*,z) p' 0 and f+('Y,z.) 1 0 < 11
and the denominator in (19) can be significantly By copaxing this ineq ality with (24) we see that thelarger than 2i . Thus it is possible to have a stable fading =emory scheme improves significantly theequilibrium even if a is very small while this is not convergence properties of the algorithm. The value ofpossible when there is a single destination. Based on B for which an optimal rate of convergence is obtainedthis fact one is tempted to conclude that the presence is the one that minimizes max fiB + (1-B) ll) -of more than one destination tends to have a benefi- i
cial effect on the dynamic behavior of the algorithm. Siiiarly one can obtain extensions of other resultsThis conjecture is supported by the results of compu- obtained Ln EI in connection with averagingtational experiments, but we have been unable to aigorithos for the single destination case.formulate it and establish it mathematically. 3. Cocoutational Results
As in (1), it is possible to define and analyze
an algorithm where the bias a depends on the current We experimented with a 30-node ring network androuting Y Similar results as those of (11 can be consiee- d a synchronous and an asynchronous algorithm
with evenly spaced delay broadcasts. A fading mmoryobtained, so we ormit the details. scheme was used to average delays corresponding to
the present and past routings. In this scheme "delays'Basing the Routing Decision on More than one Past are computed at each iteration by means of the formulaRoutings
It was seen in Ill that the dynamic behavior ofthe algorithm for the single destination case can be [New Delay of Link (i)I -
improved by employing some form of averaging of the -B[Old Delay of Link (ii)I +effects of several past routings. This analysis andthe corresponding conclusions can be generalized to + (l-0)(0.05) (Current flow of (i,U)|the multiple destination case.
As an example consider the case where the length (25)function Ik, 6 ) used in generating Yk+l depends on The scalar B is referred to as the decay factor and
all previous routings Yk' Ykl via a fading takes values in the interval 10.1). Large values ofkeory s he e OB im pjy a greater degree of averagin g with delays
memory scheme of the form corresponding to past routings. The length of link
(i,.I) used in the shortest path computation was taken6k(t) - B06klt) + (1-O)dIf(Yk't)], vte[ol to be I
+ (t) - 06- (t) + Cl-O)dlf (Yk 0) * vte[,l) [Length of Link (,)1- Bias + (New Delay Link (,fJlk k-I k' The bias depends on the current routing (c.f. (11)
and was taken to be "--
where Be(0,1) is the decay factor. Then the linearized
system corresponding to an equilibrium Y* is given Bias - 0.02 x (Sum of most recentl- "reported delayiby (cf.(lj over all links).
,(~l_ )j_ + 82 In the synchronous algorithm ll nodes "reWort"
k+l ay k+ Yk-2 ] their link delays simultaneously at each iterationWrite and all of these delays are used in the shortest path
computation that determines the new routing.
- YT + BY.. + In the asynchronous algorithm nodes comipte their... a delays at every iteration by using equation (25).
Then we have for all k mowever they "report" their link delays only everyC 30th iteration, with node I reporting at the let
Yk+l a(-0) Yk+0
(1 -0)
2 k iteration, node 2 at the 2nd iteration and so on. Thasfor exan=ple at the 31st iteration node I will report
z l i YkB " his link delays and the new routing will be competedSk0 Ii I k + kon the basis of the delays reported by node 2 at the
This system is stable if all sigenvalues of the 2nd iteration, by node 3 at the third iteration, and
matri so on. This procedure is patterned after the asyn-
1 -8 01-B) I (Y The traffic inputs for destination 30 are given at
matrix chroncus operation of he algorithm in the MAnT.
3Y the bottom of tables 1-3. The traffic sent by every
node to eath of the other destinations 1.2.29 was
I I taken to be the same and equal to A. We consider two
where I is the NOH identity matrix, lie within the values A - 0.05 and A - 1. In all runs the initialunit circle. Since this matrix has rank H it has N routirg for destination i was taken to be (I + 15)unitcirle. inc thi marix as ank ith N (r-odulo 3 1. We show in" tables I - 3 the generatedeigenvalues equal to zero. Its remaining N eigen- Wvalues can be shown to equal the eigenvalues of the sequences of successive routings for destination 30,
(.e. the s.quences of the numerically largest nodematrx BI + (1-0) 3 which are '(B)A, roIti' ; ressages to destination 30 in the Clockwise
a d rect o). The sequences of routings for the otherI,...,N where A i is the ith eigenvalue of dest rsatlons are not shown. They exhibited very
130
similar behavior. An entry a/b for the a:3' - xoncus node netvorks. Furtherore, as the level of X of traffic
algorithm means that during a 30-iteration cycle the input to the destinations I through 29 is increasedandnminimum and l&simum routirgs are a and b. The results te contribution of all destinations to the total traf-indicate that all the conclusions reached regarding fic is rmre nearly equal, th. dynamic behavior of the
the effects of bias and averaging are valid for finite algo-it-s Is i~proved.
onl:...+ .... ________l~0 .39 0 . 9 0 0 . .9 I O l 1
a. L9l It . C"i CS 0 1 3 01 1 0.0 a 0
Is o 15 1 2 Is I j /Uj Ifs12 .1/2 /10 1124 sll 124i 3li
21 2 4 1.2 2. 15 oil$ 113 9/22 1351a 101i1.3 1 U 11 W 2 13116I . ., I/ no ...
to.. " I " .. .... --...- a+I-2 0 5 0 1 s U 0 6I 0/. al2 1 0 l21 1II 11 15/1| 10/l9 n% 1 41111 10111 Il
5- W.. is.i....2 22 12 1. ae 15 Via I' l% 10/1 1 /25 ./ U 1 3 0 /
0 S 12 11 Is 11111 0/2I 1 10/ 1 1011 15001.U %0'Il I~ll3 I I 14 15 is 0/1 !Its w it 11113 Wh, iIs1 loill ISM1
I5 8 to I1 14 41 10113 is1i 10/11 k5115 1 0 I 1.513 leftl k1/13
2I I I I I liI S 1
2....... 0... . @:, .-o, , o-o 0, . , 1,0= 2.04 0 . : .. '4 7 20 5 0.9/1 03 0 1.90 :0.99 : I- 0.11
M*o.. 0 .___0"___"
L 1 0.0s I 0.e5 1 0.01 1 0.01 1 0.01 1
1 13 13 is ULF 1171 1/26 1119 112 V3/0 12/2 1/20
4 3 It 3 1I. 7114. 13111 112/2 ie1 51 3 1 311 1 12 0/ 116
14 is 24 IU 21 13 7/12 1.1/u 0/Il 11/IS ju/4 &"4 LfU/s 11M
s 1 1 1 4 I 1 4/ l S i l l3 3 4/ 1 0 1 1/t e 3 1 0 / I l 1 ,I
1 5 9 13 1t1O 2 W/.4 0/l2 14/I% 9/20 u/ 30/20 14/ U
26 2i i 14 1. $/it 1.4 IA 4 0/ 14/14 20 34/14 20 4/ILS
U U : 3 1 4 :./U3 04/14 .IOU 14/114 ern0 14/2 1U 1tt /IS
i 0 1 .4 2- 1 2.5 1& /1 .I I& 7/l. 1 U14114 1.0 /1" long -
1 1 13 it lb /16 214 11.16 14/IA4 Me "A&/ to0ile 14114
20 24 20 13 to 1& 4./14 L.214 012/412 W 1 s 14.114 1011 W14
T.41.. 2: 111-*0-021 (7-1. t.0.o). 1-1.. .- A"4- 30'- 2 ~ 0.. jl
6.001 1 '40.0 0. ,5 1 0.01 1 0.6 a 0.05 0.0 1.
I 1 13 Is I 1 112 2/1.9 I/1 21.9 1hM s/20 1127 i/l
is to Is to Is I 8/11 13/14 1.0/2 13111 / 1U I.4/Is9 15/Il 1/f1I
a 0o II i I I25 a22 3 1/16 li 13,16 11 11t4 1/14 0 1111
a? 2 4 . . 1.1 l 1.9/1. I.1 I| 1 $101.4 1 .1/1 1 21/2l4 1.lbl 130
t 1 2 1 11 1 24/14 I%11 111 1 C011 1 15 13 /4 1 " It/1.G I U I ll L iflo u1 istol IV&l$ Ili |ISl 1511 ILln
1.1. 24 22 24 14 I I 1111 0I11 1$111 2121 1 14 I91.0 116
4 2 , 1. II II, I i Il l 41/ 1.4 .9/4 1 /24 11 /Il 2!1/291 /14 ./1I I1 n IWI isi
1 AI $. I A 1 1 16 n 1 511 6tli. -7-WI
Acknowledgement:[1 IThis investigation was conducted in part at Solt n.inmize oPtf+(t)ldt + p[f'(t)lJdt
Beranek, and Newman, Inc., Cambridge, Mass. and support- Ned by ARPA Contract MDA 903-78-C-0129 and in part at subject to I riJT)uI(T)dT - f+(t) - 0the University of Illinois, Urbana, Ill., and supported i- xi t 'by Grant NSF-ENG-77-15949. The comments of JohnMcQuillan, Ira Richer and Eric Rosen, and the assist- &.a. te(0,l1.ance of Eli Gafni with the computational experimentsare gratefully acknowledged. N 4 r )u T)dT - t 0.
References: a~.tc[0,1|
(11 D.P. Bertskas, "Dynamic Models of Shortest PathRouting Algorithms for Communication Networks', +Laboratory for Information and Decision Systems 0 <u ut(t), u 1 (t)Working Paper, Massachusetts Institute of Techno-logy, Cambridge, Mass., June 1979. a.a. te[o,lj,
121 J.M. Ortega and W.C. Pheinboldt, Iterative f4* , "~ u- eL.2 m(O., i .Solution of Nonlinear Equations in SeveralVariables, Academic Press, N.Y., 1970. where a.a.te[O,l means for almost all teto.i1 with
respect to Lebesgue measure.
(3) D.G. Luenberger, Optimization by Vector SpaceMethods, J. Wiley, N.Y., 1969. To each routing Y - CyL. N
) with correspond-
ing vector Z - (z,.- . ZN) defined by (7) we can
associate the following feasible solution of problem(A.2)
Appendix A: Proof of Proposition 1 f+(t) - f (Yt). f-(t) - f (Yt) (A.3)
The function g: (0,11 N . 10,11" is clearly continuous
so by Brower's fixed theorem ((21, p. 161) it has afixed point Y* which is an equilibrium. 0 if tlRi
n order to show that Y In unique and minimizes uIt) = Mover all Yetom I if e
J (Y) - p+tf+(Yt)ldt +f1 pt f(Yt)ldt (A.l) zi'xi
we consider an associated convex programing problem. 0 if tAi"
Let L 10,11 be the space of (equivalent classes of) u1 (t) I i*xi
square integrable function on 10,1] with norm I if tea1 2 1/2
Ilfil- (fO If(t)32 dt)l/dt (A.4)
The Cartesian product LV of n copies of L2(0,1. where The corresponding value of the objective of problem
(A.2) Is J(Y). The fact that an equilibrium routingn is any positive Integer, will have the norm i . f he f olloing lm ot
n 2 1/2minimizes J(Y) follows from the following laMas
II[ .)tI ) 12
i Lema A.1: If Y is an equilibrium routing them the
functions f . f ui.u;. l-.... o ,,s9 nd-With this norm L(O,l is a Hilbert space with inner ing to Y, via equations (A.3) and (A.4) are an optimalproduct solution of problem (A.2).
<(f. f )(g .. gn > - M dt , Proof, Consider the functions dlf IT .. )).
fdf all d....Ye)L 2 (0,l. By the sufficiency theorem of
for all If 1 . ,n L:(Ol3,(g.. . gn)eL[0,1. Any (33, p. 220 the result will follow if we show that
element (g. ..... )eP10 ,1) defines a bounded linea f*+ "s
functional on L3(0,11 by means of the preceding equa- problem
tion, and all bounded linear functionals on L [,1 pm2
can be defined this way.
We consider the following convex programming
problem in j2N2
(0 ,11 which fells within the frame-2
work given in Luenberger ((33, Chapter 8)
132
minimize ptff(t)I + pit-() that (A.7) holds for all te[O,li. This together withthe assption d(O) >0 implies that , T2 . Q.E.D.N
+
dif-(Y IM I) l,[ I: M :. () 4T-f'lt)) atI Ixi.t
--i (atI + O (t uj(t) +u;)4(t) I
subject to O<u+ (t), ,t), u+(t) + u -
e.a.te 0 l, i - 1 ... , .
CA.S)
Define
bCtM - f d f+ CT* -t) dT
D: (t) - f -tAx
A straightforward calculation using integration byparts shows that problem (A.5) can be written as
minimizeO{pf+(t)] + ptf t)]
- d(f+(Y*,t)]f*Ct) - daf-C.t)3f t)
subject to O<u+(t), O_ t u * u;t) - 1.
a.a.te[O,l), i-l,....,H " A-
UsLna the fact that - d, It is easy to see tbat
I U, UL I - 1 ,..°N are an optimalsolution to this problem. Q.E.D.
It remains to show uniqueness of the equilibrium.Since d(f) > 0 for all t, all d is monotonicallyincreasing, the function p is strictly convex. LAt Fbe the convex set of all(f ,f-)eL2(0,11 for which
there exist f+. C. ul. u1 I - 1....,4 which are
feasible for problem (A.2). Problem (A.2) can bewritten as
minimszefJ (pif+ t) + ptf-lt)3)dt (A-6)
subject to (ff')e•.
If Y; and Y; are to equilibria, then
tf*(¥;" f (Yl*," ) mnd'f (Y;,' , " ;,',]
are both optimal solutions of problem (A.6). Using theetict convexity of p, it follows that except for tin a aet having Lebesgue measure zero
+ -f;.t, ,.,;.t),. ,-(,.,t) ,-Y;,t).
+ *(A.7)Since f(ty;,.), f (Y2 .'.), f CT f1 (,.', are
continuous overywhere except at x1 . xN, it follows
133